Anthropic ships Opus 4.8 and files to go public — our whole stack, one week.

1. 01

   Claude Opus 4.8 ships — and fast mode gets much cheaper in Claude Code

   May 28 release: warmer, more natural, better-calibrated thinking effort. The one that matters for us — Anthropic says fast mode for 4.8 is now much more affordable in Claude Code, the surface we run in daily. Reception is strong with an honest incremental-vs-leap debate.

2. 02

   Anthropic files a draft S-1 and raises at a $965B valuation

   Adjacent to our build, but big: a confidential SEC S-1 (Jun 1), a $65B Series H at $965B post-money, and run-rate revenue past $47B. The company we just consolidated onto is heading toward a public offering — vendor-stability context for the bet.

3. 03

   Claude Code hardens — a security plugin and a reliability pass

   Insiders shipped a security-guidance plugin that finds and fixes vulnerabilities, plus a responsiveness/reliability push. Claude Code is in our stack, so its hardening is our hardening — worth a look at the plugin.

4. 04

   Anthropic's self-hosted sandboxes land — and Gemini follows the same week

   Last edition's watch-item shipped: Managed Agents now run tool execution on your own infra while the agent loop stays managed. Gemini announced the same primitive days later — managed agents is now a category move, not one vendor's feature.

5. 05

   Next.js now ships an agent-eval harness at nextjs.org/evals

   Vercel is publicly ranking models on Next.js agent tasks. Opus and GPT-5 lead; the open model MiniMax M3 sits right behind at ~10× lower cost. A useful, neutral benchmark for anyone choosing a coding model.

Claude Opus 4.8 shipped May 28. Anthropic's framing is refinement over reinvention: it reads nuance better, feels more natural and collaborative, and lets you calibrate thinking effort more precisely. The line that matters for us is operational — fast mode for Opus 4.8 is now much more affordable in Claude Code, the surface we use every day. That's a direct cost/UX lever, not a model-card footnote.

Reception is strong but honest. Insiders called it warmer and more capable; skeptics found it incremental in daily use while conceding the gap is real on harder tasks. A separate thread claimed a benchmark top spot — treat leaderboard claims as unconfirmed until Anthropic's own post lands. For Onlook the takeaway is concrete: we're Anthropic-primary and Claude Code is in the stack, so a cheaper fast tier changes the economics of how aggressively we lean on it.

A

Alex Albert

@alexalbert__ · May 28

Excited to release Opus 4.8 today — it understands nuance better and feels much more natural. Fast mode is much more affordable now. Try it in Claude Code.

In the same window, the supplier itself moved. Anthropic confidentially submitted a draft S-1 to the SEC on June 1 — the first concrete step toward a public offering — days after disclosing a $65B Series H at a $965B post-money valuation and run-rate revenue past $47B. This is adjacent to what we build, not a tool change. But it's not noise either: it's vendor-stability context for the bet we just made consolidating onto Anthropic.

The honest read cuts both ways. A supplier at public-company scale is durable, well-capitalized, and unlikely to disappear — good for a small team betting its core loop on one provider. The flip side is that a soon-to-be-public company optimizes for margin and predictability, which is exactly the pressure behind this window's other Anthropic story — metered, usage-based billing. Worth a periodic glance, not a standing worry.

Two Claude Code improvements surfaced from the team building it. A security-guidance plugin that identifies and helps fix vulnerabilities, paired with an engineering post on how agent access and permissions should evolve; and a separate push on responsiveness and reliability. Neither is a headline release, but Claude Code is the surface our work runs through — its hardening is ours.

B

Boris Cherny

@bcherny · May 26

Shipped a security plugin for Claude Code — it identifies and helps fix vulnerabilities in your code.

Edition 07's watch-item landed with a date. Anthropic's Managed Agents can now run tool execution inside self-hosted sandboxes on your own infra (Cloudflare, Daytona, Modal, Vercel) with MCP tunnels, while the agent loop stays managed by Anthropic. The strategic shape we flagged holds: own the execution environment, keep the orchestration off your plate.

The new signal is competitive. Within days, Gemini announced managed agents in its API too — so this is a category move, not a single vendor's feature. For us it strengthens the pattern worth prototyping if we ever execute generated code: managed loop, our sandbox, our data boundary.

D

@DhravyaShah · May 25

Following Anthropic, Gemini has also launched managed agents in the API.

A genuinely non-Anthropic item, and a useful one. Next.js now publishes an agent-eval harness at nextjs.org/evals — ranking models on real Next.js agent tasks. Opus and GPT-5 lead; the open model MiniMax M3 sits right behind at roughly 10× lower cost. For anyone choosing a coding model, that's a neutral, framework-native benchmark rather than a vendor's own scorecard.

R

Guillermo Rauch

@rauchg · Jun 1

MiniMax M3 is now the leading open model on the Next.js agent evals — right behind Opus & GPT-5, but ~10× cheaper.